Gemini 3.1 Pro (Preview)

The display name ends in "(Preview)," and that's the honest label: Gemini 3.1 Pro is the flagship-tier model Google gates behind a preview window, and this node is your way to drive it from inside a ComfyUI graph. You get a proper LLM step - prompt expansion, multi-step reasoning about what a prompt will produce, structured rewriting - with the quality of the Pro line and none of the local cost. It's Google's model, billed through your BizyAir key, running on their hardware.

The "Pro" positioning matters for one practical reason: this is the node you swap in when Flash-class output stops being good enough. If your prompt pipeline occasionally produces something the text model clearly misjudged, Pro is the upgrade to try. It's slower and pricier than Flash-Lite - that's the deal with a flagship.

Inputs and outputs

• system_prompt - required. The role and constraints.
• user_prompt - required. The task.
• temperature - 0 to 2, default 1.
• max_tokens - default 32768, up to 65536.
• enable_thinking - default false. This is the interesting one: Pro is built to reason, but the node ships with thinking off. For serious prompt engineering work, flip it on and watch the quality jump - at the cost of latency.
• enable_search - default false here, unlike the Flash models. This is a deliberate default: Pro is meant for pure reasoning, not web-fetching. Turn search on only if your task genuinely needs current information.

Outputs: result (STRING) and request_id. Remember a text preview node if you want to see the output in the graph.

How it works

Same pack machinery as the rest of the Gemini line: serialize the prompts and sampling settings, hit the gemini-3.1-pro-official/large-language-models endpoint, poll for the completion, return the text. api_config overrides base URL and key per request; skip_error swaps a crash for a placeholder on failure.

The honest take

Here's where to spend the difference. The typical ComfyUI user doesn't need Pro for prompt expansion - Flash-Lite covers that. Pro earns its credits when you're doing something judgment-heavy: planning a multi-step workflow, reasoning about why a prompt failed, or acting as the "editor" in a generate-then-critique loop where the critique has to be sharper than the generator. And since thinking is off by default, you can cheaply run it like a Flash model until you actually need the reasoning - then flip it on and let the flagship do what it's for.

Install

With the pack:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAirPlus.git
cd BizyAirPlus
python -m pip install -r requirements.txt

Restart ComfyUI, confirm BizyAirPlus is ON and keyed. Start with thinking off, and turn it on the first time the output is too shallow. Preview or not, it's the sharpest text node in the pack.